The high precision of the tensor on the mask tensor causes errors in the Expand Mask results

[rubi-du]

When there is a minimum value in the mask tensor, it can lead to incorrect Expand mask results and loss of mask calculation

• The mask tensor like this：

• The result as follow：

As shown in the above figure, directly converting the mask to an image can work normally, but the mask is lost after expanding it

[Acly]

Yes, "grow" only works for binary masks (all values 0 or 1).

A more general dilation for smooth mask is a lot more expensive I think (at least I can't think of a method to do it as a convolution).